Research Home Topics EU AI Act Countries Sources Calendar Methodology Authors

General-Purpose AI Model Obligations Under the EU AI Act (Articles 51, 53, and 55)

A separate track from the high-risk-system rules covered elsewhere in this pillar: every general-purpose AI model provider faces documentation and transparency duties from Article 53, models crossing a 1025 FLOP compute threshold face additional systemic-risk obligations under Article 55 — and this entire tier has applied since 2 August 2025, a full year ahead of the Act's general application date.

Key context
"General-purpose AI models that could pose systemic risks must undergo adversarial testing, report serious incidents to the AI Office, and implement cybersecurity measures at a higher standard."
EU AI Act, Article 55
A bank of server racks in a data centre with status lights, evoking the compute-threshold test that determines whether a general-purpose AI model carries systemic risk under Article 51.

Builds on AR-000200 (The EU AI Act Explained), AR-000210 (High-Risk AI Systems), and AR-000230 (Transparency). AR-000200 introduced GPAI briefly as one of the Act's tiers; this article supplies the Articles 51/53/55 detail it deferred, and independently confirms (rather than repeats) AR-000200's application-date claim.

Executive summary

What this means

Who this applies to

Regulatory context

Requirements

What organizations must do

  • Treat the 1025 FLOP threshold as a planning input, not a retrospective check — the two-week notification clock starts running once the threshold "becomes known" to be met, not only once it is actually met.
  • Confirm all of the Article 53(2) open-source exemption's conditions are met simultaneously before relying on it, and remember it disappears entirely for systemic-risk models.
  • Treat Article 53 compliance as a live obligation, not a future one — Chapter V has applied since 2 August 2025.
  • Request Article 53(1)(b) Annex XII information directly from an upstream model provider before building a downstream system on top of it.

Common mistakes

Assuming GPAI obligations only apply to a small number of well-known frontier-model providers is a common mistake — Article 53's baseline obligations apply to providers of general-purpose AI models as a category, with systemic risk under Article 55 as an additional tier for a narrower subset, not a replacement trigger.

Treating the open-source exemption as a blanket carve-out for any freely available model is a second mistake — the exemption has multiple cumulative conditions and is unavailable entirely for systemic-risk models, regardless of licensing.

Assuming the GPAI tier's earlier 2 August 2025 application date means its fine regime is equally mature is a third mistake this article specifically corrects — Article 101 was excepted from that date and instead applies from the Act's general 2 August 2026 date, a full year later.

Confusing a general-purpose AI model's own Chapter V compliance with the Chapter III high-risk classification of a downstream AI system built on top of it is a fourth mistake — the two are separate regimes, and satisfying one does not satisfy the other.

Open questions

  • Whether Chapter V was genuinely unamended by the Digital Omnibus, given the consistent absence of disclaimers across Articles 51, 52, 53, and 55.
  • What the Commission's published guidelines on the scope of GPAI obligations substantively say — not reviewed on this pass, only their existence was noted.
  • The exact mechanics of Article 52's Commission-designation and reassessment procedure beyond paragraph 1's notification duty — corroborated via a secondary source only, not independently re-verified word-for-word.

Frequently asked

FAQ
Did the Digital Omnibus amend any of Articles 51 through 56, given the absence of a disclaimer on their Service Desk pages?
Not independently verified — the consistent absence across all of Chapter V is noted but not treated as confirmation either way.
FAQ
What do the Commission's published guidelines on the scope of GPAI obligations actually say?
Not reviewed on this research pass; only the existence and publication date of the guidelines document were noted.
FAQ
What is the Commission's exact procedure for designating a model as carrying systemic risk absent the compute threshold, and how does the six-month reassessment window work in practice?
Corroborated via a secondary source but not independently re-verified word-for-word against the primary Article 52 text beyond paragraph 1.

Go deeper

Apply this in practice

AGNUS Academy turns AI governance research into hands-on skills — exercises, case studies, and frameworks built on the research you just read.

Explore the Academy →